[data]
type=data
dataIdx=0

[labels]
type=data
dataIdx=1

[conv1a]
type=conv
inputs=data
channels=3
filters=48
padding=0
stride=4
filterSize=11
initW=0.01
partialSum=11
sharedBiases=1
gpu=0

[conv1b]
type=conv
inputs=data
channels=3
filters=48
padding=0
stride=4
filterSize=11
initW=0.01
partialSum=11
sharedBiases=1
gpu=1

[conv1c]
type=conv
inputs=data
channels=3
filters=48
padding=0
stride=4
filterSize=11
initW=0.01
partialSum=11
sharedBiases=1
gpu=2

[conv1d]
type=conv
inputs=data
channels=3
filters=48
padding=0
stride=4
filterSize=11
initW=0.01
partialSum=11
sharedBiases=1
gpu=3

[rnorm1a]
type=cmrnorm
inputs=conv1a
channels=48
size=5

[rnorm1b]
type=cmrnorm
inputs=conv1b
channels=48
size=5

[rnorm1c]
type=cmrnorm
inputs=conv1c
channels=48
size=5

[rnorm1d]
type=cmrnorm
inputs=conv1d
channels=48
size=5

[pool1a]
type=pool
pool=max
inputs=rnorm1a
sizeX=3
stride=2
channels=48
neuron=relu

[pool1b]
type=pool
pool=max
inputs=rnorm1b
sizeX=3
stride=2
channels=48
neuron=relu

[pool1c]
type=pool
pool=max
inputs=rnorm1c
sizeX=3
stride=2
channels=48
neuron=relu

[pool1d]
type=pool
pool=max
inputs=rnorm1d
sizeX=3
stride=2
channels=48
neuron=relu

[conv2a]
type=conv
inputs=pool1a
filters=128
padding=2
stride=1
filterSize=5
channels=48
initW=0.01
initB=1
partialSum=9
sharedBiases=1
neuron=relu

[conv2b]
type=conv
inputs=pool1b
filters=128
padding=2
stride=1
filterSize=5
channels=48
initW=0.01
initB=1
partialSum=9
sharedBiases=1
neuron=relu

[conv2c]
type=conv
inputs=pool1c
filters=128
padding=2
stride=1
filterSize=5
channels=48
initW=0.01
initB=1
partialSum=9
sharedBiases=1
neuron=relu

[conv2d]
type=conv
inputs=pool1d
filters=128
padding=2
stride=1
filterSize=5
channels=48
initW=0.01
initB=1
partialSum=9
sharedBiases=1
neuron=relu

[rnorm2a]
type=cmrnorm
inputs=conv2a
channels=128
size=5

[rnorm2b]
type=cmrnorm
inputs=conv2b
channels=128
size=5

[rnorm2c]
type=cmrnorm
inputs=conv2c
channels=128
size=5

[rnorm2d]
type=cmrnorm
inputs=conv2d
channels=128
size=5

[cnorm2a]
type=rnorm
inputs=rnorm2a
channels=128
size=5

[cnorm2b]
type=rnorm
inputs=rnorm2b
channels=128
size=5

[cnorm2c]
type=rnorm
inputs=rnorm2c
channels=128
size=5

[cnorm2d]
type=rnorm
inputs=rnorm2d
channels=128
size=5

[pool2a]
type=pool
pool=max
inputs=cnorm2a
sizeX=3
stride=2
channels=128

[pool2b]
type=pool
pool=max
inputs=cnorm2b
sizeX=3
stride=2
channels=128

[pool2c]
type=pool
pool=max
inputs=cnorm2c
sizeX=3
stride=2
channels=128

[pool2d]
type=pool
pool=max
inputs=cnorm2d
sizeX=3
stride=2
channels=128

[conv3a]
type=conv
inputs=pool2a,pool2b,pool2c
filters=192,192,192
padding=1,1,1
stride=1,1,1
filterSize=3,3,3
channels=128,128,128
initW=0.03,0.03,0.03
partialSum=13
sharedBiases=1
neuron=relu
gpu=0

[conv3b]
type=conv
inputs=pool2a,pool2b,pool2d
filters=192,192,192
padding=1,1,1
stride=1,1,1
filterSize=3,3,3
channels=128,128,128
initW=0.03,0.03,0.03
partialSum=13
sharedBiases=1
neuron=relu
gpu=1

[conv3c]
type=conv
inputs=pool2c,pool2d,pool2a
filters=192,192,192
padding=1,1,1
stride=1,1,1
filterSize=3,3,3
channels=128,128,128
initW=0.03,0.03,0.03
partialSum=13
sharedBiases=1
neuron=relu
gpu=2

[conv3d]
type=conv
inputs=pool2c,pool2d,pool2b
filters=192,192,192
padding=1,1,1
stride=1,1,1
filterSize=3,3,3
channels=128,128,128
initW=0.03,0.03,0.03
partialSum=13
sharedBiases=1
neuron=relu
gpu=3

[conv4a]
type=conv
inputs=conv3a
filters=192
padding=1
stride=1
filterSize=3
channels=192
neuron=relu
initW=0.03
initB=1
partialSum=13
sharedBiases=1

[conv4b]
type=conv
inputs=conv3b
filters=192
padding=1
stride=1
filterSize=3
channels=192
neuron=relu
initW=0.03
initB=1
partialSum=13
sharedBiases=1

[conv4c]
type=conv
inputs=conv3c
filters=192
padding=1
stride=1
filterSize=3
channels=192
neuron=relu
initW=0.03
initB=1
partialSum=13
sharedBiases=1

[conv4d]
type=conv
inputs=conv3d
filters=192
padding=1
stride=1
filterSize=3
channels=192
neuron=relu
initW=0.03
initB=1
partialSum=13
sharedBiases=1

[conv5a]
type=conv
inputs=conv4a
filters=128
padding=1
stride=1
filterSize=3
channels=192
initW=0.03
initB=1
partialSum=13
groups=1

[conv5b]
type=conv
inputs=conv4b
filters=128
padding=1
stride=1
filterSize=3
channels=192
initW=0.03
initB=1
partialSum=13
groups=1

[conv5c]
type=conv
inputs=conv4c
filters=128
padding=1
stride=1
filterSize=3
channels=192
initW=0.03
initB=1
partialSum=13
groups=1

[conv5d]
type=conv
inputs=conv4d
filters=128
padding=1
stride=1
filterSize=3
channels=192
initW=0.03
initB=1
partialSum=13
groups=1

[pool3a]
type=pool
pool=max
inputs=conv5a
sizeX=3
stride=2
channels=128
neuron=relu

[pool3b]
type=pool
pool=max
inputs=conv5b
sizeX=3
stride=2
channels=128
neuron=relu

[pool3c]
type=pool
pool=max
inputs=conv5c
sizeX=3
stride=2
channels=128
neuron=relu

[pool3d]
type=pool
pool=max
inputs=conv5d
sizeX=3
stride=2
channels=128
neuron=relu

[conv6a]
type=conv
inputs=pool3a,pool3b,pool3c
filters=128,128,128
padding=1,1,1
stride=1,1,1
filterSize=3,3,3
channels=128,128,128
initW=0.03,0.03,0.03
initB=1
partialSum=4
neuron=relu
gpu=0

[conv6b]
type=conv
inputs=pool3a,pool3b,pool3d
filters=128,128,128
padding=1,1,1
stride=1,1,1
filterSize=3,3,3
channels=128,128,128
initW=0.03,0.03,0.03
initB=1
partialSum=4
neuron=relu
gpu=1

[conv6c]
type=conv
inputs=pool3c,pool3d,pool3a
filters=128,128,128
padding=1,1,1
stride=1,1,1
filterSize=3,3,3
channels=128,128,128
initW=0.03,0.03,0.03
initB=1
partialSum=4
neuron=relu
gpu=2

[conv6d]
type=conv
inputs=pool3c,pool3d,pool3b
filters=128,128,128
padding=1,1,1
stride=1,1,1
filterSize=3,3,3
channels=128,128,128
initW=0.03,0.03,0.03
initB=1
partialSum=4
neuron=relu
gpu=3

[fc1024a]
type=fc
inputs=conv6a,conv6b,conv6c
outputs=1024
initW=0.01,0.01,0.01
initB=1
neuron=relu
gpu=0

[fc1024b]
type=fc
inputs=conv6a,conv6b,conv6d
outputs=1024
initW=0.01,0.01,0.01
initB=1
neuron=relu
gpu=1

[fc1024c]
type=fc
inputs=conv6c,conv6d,conv6a
outputs=1024
initW=0.01,0.01,0.01
initB=1
neuron=relu
gpu=2

[fc1024d]
type=fc
inputs=conv6c,conv6d,conv6b
outputs=1024
initW=0.01,0.01,0.01
initB=1
neuron=relu
gpu=3

[hs1a]
type=hs
keep=0.5
inputs=fc1024a

[hs1b]
type=hs
keep=0.5
inputs=fc1024b

[hs1c]
type=hs
keep=0.5
inputs=fc1024c

[hs1d]
type=hs
keep=0.5
inputs=fc1024d

[fc1024ba]
type=fc
inputs=hs1a,hs1b,hs1c
outputs=1024
initW=0.01,0.01,0.01
initB=1
neuron=relu
gpu=0

[fc1024bb]
type=fc
inputs=hs1b,hs1a,hs1d
outputs=1024
initW=0.01,0.01,0.01
initB=1
neuron=relu
gpu=1

[fc1024bc]
type=fc
inputs=hs1c,hs1d,hs1a
outputs=1024
initW=0.01,0.01,0.01
initB=1
neuron=relu
gpu=2

[fc1024bd]
type=fc
inputs=hs1c,hs1d,hs1b
outputs=1024
initW=0.01,0.01,0.01
initB=1
neuron=relu
gpu=3

[hs2a]
type=hs
keep=0.5
inputs=fc1024ba

[hs2b]
type=hs
keep=0.5
inputs=fc1024bb

[hs2c]
type=hs
keep=0.5
inputs=fc1024bc

[hs2d]
type=hs
keep=0.5
inputs=fc1024bd

[fc1000a]
type=fc
outputs=2546
inputs=hs2a,hs2b,hs2c,hs2d
initW=0.01,0.01,0.01,0.01
gpu=0

[fc1000b]
type=fc
outputs=2546
inputs=hs2a,hs2b,hs2c,hs2d
initW=0.01,0.01,0.01,0.01
gpu=1

[fc1000c]
type=fc
outputs=2546
inputs=hs2a,hs2b,hs2c,hs2d
initW=0.01,0.01,0.01,0.01
gpu=2

[fc1000d]
type=fc
outputs=2546
inputs=hs2a,hs2b,hs2c,hs2d
initW=0.01,0.01,0.01,0.01
gpu=3

[concat]
type=concat
inputs=fc1000a,fc1000b,fc1000c,fc1000d

[probs]
type=softmax
inputs=concat
gpu=0

[logprob]
type=cost.logreg
inputs=labels,probs
gpu=0